Sloppiness around failure handling of parsePGArray in pg_dump

Hi all,

Following the report of Coverity that led to 3636efa, I have reviewed
the existing callers of parsePGArray() in pg_dump and some of its
error handling is a bit sloppy.

It could theoretically be possible to reach an OOM in parsePGArray()
with a dump able to finish. This is very unlikely going to matter in
practice as an OOM when parsing an array is most likely going to
trigger a fatal failure in one of the follow-up allocations, but if
the dump is able to go through we could finish with a valid dump that
lacks some information:
- Statistics for indexes.
- Run-time configuration of functions.
- Configuration of extensions.
- Publication list for a subscription.

I would like to propose the attached to tighten the error handling in
the area, generating a fatal error if an array cannot be parsed. I
did not see the point of changing the assumptions we use for the
parsing of function args or such when it comes to pre-8.4 dumps. This
issue is unlikely going to matter in practice, so I don't propose a
backpatch.

Thoughts?
--
Michael